Magnetic Resonance Imaging Characteristics Predict Epidermal ...

Imaging, Diagnosis, Prognosis

Magnetic Resonance Imaging Characteristics Predict Epidermal Growth Factor Receptor Amplification Status in Glioblastoma Manish Aghi,1 Paola Gaviani,2,5 John W. Henson,2,3 Tracy T. Batchelor,2,6 David N. Louis,1,4,6 and Fred G. Barker II1

Abstract

Purpose:Two clinical-molecular glioblastoma subtypes have been described: ‘‘primary’’ glioblastomas arise de novo in older patients and often overexpress epidermal growth factor receptor (EGFR); ‘‘secondary’’ glioblastomas progress from lower-grade tumors in younger patients and commonly have TP53 mutations. EGFR overexpression correlates in experimental gliomas with increased angiogenesis, edema, and invasion. No radiographic predictors of molecular glioblastoma subtype are known. Experimental Design: We retrospectively reviewed 75 glioblastomas, classified as TP53mutated (n = 11), EGFR-amplified (n = 31), or neither (non-TP53/non-EGFR; n = 33). Four variables were derived from preoperative magnetic resonance imaging: (a) T2/T1, the ratio of T2-bright volume to enclosed T1-enhancing volume; (b) percentage of tumor volume that was necrosis; and (c and d) T1 and T2 border sharpness coefficients (BSC), the rates of change in grayscale intensity of adjacent 0.02-cm2 voxels traversing the anterior, posterior, and lateral borders onT1-enhanced and T2 images. Results and Conclusions: Mean T2/T1 was 4.7 for EGFR-amplified glioblastomas, greater than that of TP53 -mutated glioblastomas (2.3) or non-TP53 /non-EGFR glioblastomas (2.6; P < 0.00005). All four tumors with T2/T1 > 7.2 were EGFR-amplified; 0 of 15 with T2/T1 < 4.7 underwent gross total resection. The meanT2 BSC of EGFR-amplified glioblastomas was 33.7, less sharp (P < 0.0000005) than TP53-mutated (72.2) and non-TP53/non-EGFR glioblastomas (81.2). All 15 glioblastomas withT2 BSC < 30.8 were EGFR-amplified. Percentage necrosis and T1BSC did not differ between glioblastoma subtypes. The increased T2/T1ratio and decreased T2 BSC in EGFR-overexpressing tumors are the first radiographic distinctions described between glioblastoma molecular subtypes.These findings may reflect increased angiogenesis, edema, and/or invasion in EGFR-overexpressing tumors.

Despite aggressive management with surgery, radiation, and chemotherapy, the prognosis for patients with glioblastoma remains poor. Median survival is typically 0.05 by two-sample t test using the number of pixels in each voxel, the average grayscale intensity of the 4 to 12 pixels in each voxel, and the SD of the grayscale units of the pixels in each voxel) to the same space in the contralateral hemisphere; the average grayscale units contained within each of the voxels was then recorded; (b) the slope of a linear regression correlating average grayscale unit with voxel number (one to four) was calculated; (c) the average of the four slopes at the anterior, posterior, right, and left tumor borders, excluding borders that abutted ventricles or cortical surfaces, was the BSC, with higher BSC signifying a sharp border and lower BSC a fuzzy border (Fig. 2). Calculation of each BSC took f5 minutes per tumor. To confirm interobserver reproducibility of T1 and T2 BSCs, a second observer, also blinded to the results of molecular analysis, derived T1 and T2 BSCs for each tumor. The second observer verified that T1 BSC did not differ between glioblastoma subtypes whereas T2 BSC was less in EGFR-amplified than in non-EGFRamplified glioblastomas (P < 0.02). j statistics, derived by stratifying the BSCs of each observer into quintiles, were 0.4 for T1 BSC and 0.5 for T2 BSC, confirming significant interobserver reproducibility (P < 0.00001). Statistical analysis. S-plus software was used to derive j statistics to assess interobserver variability and receiver operating characteristic curves for T2/T1 and T2 BSC analysis. The log-rank test was used to

Fig. 1. Method of calculatingT2/T1ratio. Vitrea 2 computer software was used to derive three-dimensional volumes of T2 brightness and enclosed T1-enhancing volume (including internal necrosis). The ratio of these volumes was theT2/T1ratio. The software outlines the areas of abnormality on each axial image and the user can modify any incorrectly labeled areas. The volume is then derived by the program using the total areas of abnormalities and the thickness of each axial image.Two representativeT2 (left two images in A and B) and T1-enhanced (right two images in A and B) axial images from two tumors (A and B), one of which (A) had a lowT2/T1ratio and proved to be non-EGFR-amplified whereas the other (B) had a highT2/T1ratio and proved to be EGFR-amplified.

www.aacrjournals.org

8601

Clin Cancer Res 2005;11(24) December 15, 2005


Fig. 2. Method of calculating BSC. To calculateT2 BSC, four adjacent 0.02-cm2 voxels were placed along a line perpendicular to the anterior, posterior, and lateral tumor borders on the axial image containing the largest T2-bright area. Axial T2 images with the largest area of abnormality from two different tumors (A and B). The posteriorT2 bright borders are enlarged. The first voxel is placed as close to the edge of T2 brightness as possible such that its grayscale units do not deviate from that of a voxel in the same position in the contralateral hemisphere.This first voxel is followed by three more adjacent voxels running posterior to anterior. The slope of a linear regression connecting the grayscale units of the four voxels is derived. The process is repeated for the three other borders (except for borders that lie on cortical surfaces or against ventricles) and theT2 BSC is the average of the slopes at each border. The tumor shown in (A) had the largest T2 BSC, signifying a sharpT2 bright border, and proved to be non-EGFR-amplified, whereas the tumor in (B) had the smallestT2 BSC, signifying a fuzzyT2 bright border, and proved to be EGFR-amplified. ForT1BSC, three adjacent voxels were used on the axial image containing the largest gadolinium-enhanced area.

compare Kaplan-Meier survival curves of glioblastoma subgroups. Spearman’s rank correlation coefficient was used to quantify the correlation between ordinal data such as age, T2/T1, T1-enhancing volume, T2 BSC, and percent necrosis. Student’s t test was used as a parametric test to compare the differences in the radiographic variables (T1 BSC, percent necrosis) between glioblastoma subtypes. ANOVA was used to compare the features of tumors in each of the four lobes. For nonparametric comparisons (T2/T1, T2 BSC), the Kruskal-Wallis test was used to compare multiple groups and the Mann-Whitney U test was used to compare two groups.

Results Basic MRI findings of the molecular glioblastoma subtypes. There were 31 EGFR-amplified, 11 TP53-mutated, and 33 non-TP53/non-EGFR glioblastomas with mean T1-enhancing volumes (including internal necrosis) of 31.2, 30.9, and 44.0 cm3, respectively (P > 0.05). These tumors were then analyzed for four radiographic variables that could be derived from their T1 and T2 studies—T2/T1 ratio, percent necrosis, and T1 and T2 BSCs. T2/T1 ratio of molecular glioblastoma subtypes. T2-bright tissue volumes at the margins of malignant gliomas reflect edema and tumor invasion (10). We compared the T2/T1 ratio, the ratio of T2-bright volume to enclosed T1-enhancing volume (Fig. 1), between molecular glioblastoma multiforme subtypes. T2/T1 ratio was significantly greater in EGFR -amplified glioblastomas than in the other groups (P < 0.00005, MannWhitney U test). Mean T2/T1 was 4.7 for EGFR-amplified glioblastomas (range 1.3-15), 2.3 for TP53-mutated glioblastomas (1.0-4.8), and 2.6 for non-TP53/non-EGFR glioblastomas (1.0-7.2; Fig. 3A). All four tumors with T2/T1 > 7.2 (5% of the tumors analyzed) were EGFR-amplified. T2/T1 decreased as T1 volume increased and increased with the patient’s age, but neither tendency was statistically significant (P > 0.05, Spearman’s rank correlation). Nine of


44 non-EGFR-amplified glioblastomas (20.5%) underwent gross total resection, exceeding the 4 of 31 EGFR-amplified glioblastomas (12.9%) that received gross total resection (P = 0.37). None of 15 tumors (20% of the tumors analyzed) with T2/T1 > 4.7 underwent gross total resection. T2/T1 did not correlate with survival (P > 0.05)—the median survival of patients with T2/T1 > 3 (the median for all glioblastomas in this report) was 362 days, compared with 385 days for patients with T2/T1 < 3 (P > 0.05). Average T2/T1 values were 4.9, 3.7, 3.0, and 2.7 for lesions involving the parietal, temporal, occipital, and frontal lobes, respectively (P = 0.056 for global comparison of the four groups by ANOVA; P = 0.007 for parietal versus frontal, P = 0.03 for parietal versus occipital, and P > 0.05 for all other comparisons). Percent necrosis found in molecular glioblastoma subtypes. The exact mechanisms of glioblastoma necrosis have not been identified. To determine whether EGFR amplification or TP53 mutations are associated with glioblastoma necrosis, we compared the percentage of radiographic necrosis found in the molecular glioblastoma subtypes by calculating the percentage of total T1 tumor volume that had a hypodense appearance contained within an enhancing border. The percent necrosis did not vary with glioblastoma subtype, averaging 29.7%, 24.0%, and 26.2% for EGFR-amplified (range, 0-80%), TP53mutated (range, 0-74.4%), and non-TP53/non-EGFR (range, 0-71.3%) glioblastomas, respectively. Necrosis also did not correlate with survival—the median survival of patients whose tumors exhibited >22.5% necrosis (the median necrosis) was 365 days, virtually identical to the 364-day median survival of patients with tumors exhibiting 0.05). T1 BSC also did not affect survival—the median survival of patients whose tumors had T1 BSCs < 99.1 (the median T1 BSC) was 365 days, comparable to the 364-day median survival of patients whose tumors had T1 BSCs z 99.1. T2 BSC of the molecular glioblastoma subtypes. Similarly, the T2 BSC was derived using the axial T2 image containing the largest T2-bright area (Fig. 2). The mean T2 BSC of EGFRamplified glioblastomas was 33.7, significantly less than the values for TP53-mutated (72.2) and non-TP53/non-EGFR glioblastomas (81.2; P < 0.0000005, Mann-Whitney U test; Fig. 3B). All 15 glioblastomas with T2 BSC < 30.8 (20% of tumors analyzed) were EGFR-amplified. T2 BSC did not correlate with survival—patients whose glioblastomas had T2 BSCs < 57.4 (the median T2 BSC) had a median survival of 345 days, comparable to the 385-day median survival of patients whose glioblastomas had T2 BSCs > 57.4 (P > 0.05). T2 BSC also did not correlate with anatomic location of glioblastoma— the mean T2 BSCs for tumors involving the parietal, temporal, occipital, and frontal lobes were 48.4, 58.7, 60.0, and 61.3, respectively (P = 0.5, ANOVA). T2 BSC decreased with increasing T2/T1 ratio in a significant manner (P = 0.048). Although T2 BSC tended to decrease with increasing patient age, the tendency was not significant (P = 0.2). Receiver operating characteristic curves for T2/T1 and T2 BSC. Receiver operating characteristic curves, which depict the tradeoff between sensitivity and specificity, were generated for T2/T1 and T2 BSC in this cohort (Fig. 4). Areas under receiver operating characteristic curves can range from 0.5, when a test has 50% sensitivity and 50% specificity and is no better than random chance, to 1.0 for a test with perfect discrimination because it achieves both 100% sensitivity and 100% specificity. The area under the T2/T1 receiver operating characteristic curve was 0.78 whereas the area under the T2 BSC receiver operating characteristic curve was 0.92 (Fig. 4), confirming the degree to which both variables balanced sensitivity and specificity, T2 BSC > T2/T1.

www.aacrjournals.org

Discussion Several phenotypic differences between primary and secondary glioblastoma subtypes have been previously described. First, there are differences in the patient populations, with EGFRamplified glioblastomas affecting an older population and more often representing de novo tumors (2). Second, there are biological differences, with EGFR-amplified glioblastomas exhibiting increased angiogenesis, likely due to increased secretion of vascular endothelial growth factor (13), and increased brain invasion, likely due to increased secretion of matrix metalloproteinases (14). Third, there are differences in outcome, with EGFR-amplified glioblastomas being potentially more resistant to radiation (3), undergoing more rapid postoperative regrowth (4), and displaying improved survival in older patients but worse survival in younger patients if TP53 is not mutated (15). Here, we report the first findings of radiographic differences between molecular glioblastoma subtypes: increased T2/T1 ratio and decreased T2 BSC (fuzzier borders) in tumors overexpressing EGFR. T2-bright tissue volume surrounding a

Fig. 4. Receiver operating characteristic curves forT2/T1and T2 BSC. Receiver operating characteristic curves plotting sensitivity as a function of the false positive ratio (1Fspecificity) are shown forT2/T1 (left) and T2 BSC (right). The areas under the curves, which can range from 0.5 if a test has no discrimination and is no better than random chance to 1.0 for a test with perfect discrimination (100% sensitivity and 100% specificity), were 0.78 (T2/T1) and 0.92 (T2 BSC).

8603



T1-enhancing tumor has been attributed to edema, microscopic infiltrative tumor, or both (10). EGFR-amplified glioblastomas could exhibit increased surrounding edema because EGFR activation enhances synthesis of hypoxia-inducible transcription factor (16), which in turn promotes increased secretion of cytokines, such as vascular endothelial growth factor, which act as mitogens for endothelial cells and act directly on endothelial cells to increase permeability selectively to proteins (13). T2-bright signal around tumors may also be explained by microscopic tumor infiltration. In T2-bright white matter beyond tumor margins, as defined by T1 contrast-enhanced MRI, nests of tumor can be identified in stereotactic biopsy specimens (10). Glioma invasion into surrounding white matter is a complex multistep process, likely starting with glioma lysing the extracellular matrix that normally limits cellular migration. Matrix metalloproteinases secreted by gliomas may aid invasion by degrading a broad range of extracellular matrix components surrounding gliomas. Matrix metalloproteinases 1 and 13 are collagenases induced by EGFR signaling and implicated in glioma cell invasion (14). Fibroblast activation protein a, also known as seprase, is an integral membrane serine protease implicated in invasion of which expression is also up-regulated by EGFR signaling (5). In addition, we have shown that there seems to be a microenvironmental selection pressure to retain double minute chromosomal amplicons of the EGFR gene at the invasive edges of glioblastomas (9). The finding of an increased T2/T1 ratio in tumors overexpressing EGFR likely reflects both the increased angiogenesis and edema caused by increased vascular endothelial growth factor secretion and the invasion by these tumors caused by increased expression of matrix metalloproteinases and other proteases. The finding of a decreased T2 BSC in tumors overexpressing EGFR, suggesting less sharpness in the T2 border of these tumors, may reflect the increased invasiveness of these lesions. Another example in which MRI findings correlated with molecular mutations of a tumor was a report documenting that allelic loss of chromosomal arms 1p and 19q in oligodendrogliomas was associated with an indistinct border on T1 images and mixed signal intensity on T1 and T2 images (17). This correlation between tumor molecular analysis and MRI was based on subjective MRI scores rather than on quantitative measures such as the ones we used. The correlation between molecular analysis and MRI described here is the first such correlation for glioblastoma and is based on objective criteria derived from computer-based image analysis that can be done in a reproducible manner by a single person rather than on a subjective review by a panel of experts. Initial studies correlating MRI findings with histologic features of astrocytic tumors focused on predicting WHO grade, which is of importance because incomplete biopsy sampling of heterogeneous astrocytic tumors could lead to inaccurate grading. Several studies found that MRIs of glioblastomas showed more hemorrhage, cyst formation, necrosis, and contrast enhancement than those of low-grade and anaplastic astrocytomas (18, 19). The present report suggests that, with further volumetric analysis, a preoperative MRI may also enable prediction of molecular glioblastoma subtype. This is the first reported use of the T2/T1 ratio for MRI analysis. The closest previously published quantitative MRI


variable was the vasogenic edema ratio which measured vasogenic edema in hematomas by taking the maximum width of edema outside hematoma on T2 and dividing by the diameter of the hematoma (20). As the vasogenic edema ratio increased, the positive predictive value that there was neoplasm inside the hemorrhage increased. Thus, the zone of T2 brightness on MRI, reflecting edema, is larger around brain tumors than around hematomas. This report expands our understanding of T2 brightness associated with glioblastomas by analyzing three-dimensional volumes rather than twodimensional diameters and by correlating findings with molecular glioblastoma subtype. Radiographic necrosis did not vary with glioblastoma subtype and the average necrosis in our glioblastomas was 27%, comparable to a prior report documenting 40% radiographic necrosis on glioblastoma MRIs (21). Glioblastoma radiographic necrosis on MRIs graded on a four-point scale has been reported to inversely correlate with patient survival in a cohort slightly smaller than ours (22) and the presence of any radiographic necrosis represented a negative prognostic factor in a multivariate analysis of over 400 glioblastoma patients in a larger series (23). However, in our patient cohort there was no effect of necrosis on survival, even when patients with no radiographic necrosis were compared with patients with radiographic necrosis. The mechanism of glioblastoma necrosis has not been identified. The lack of correlation between radiographic necrosis and glioblastoma subtype may reflect multiple effects in EGFR-amplified glioblastomas that counteract each other. On the one hand, vascular endothelial growth factor – mediated increases in endothelial thromboplastin activity could promote necrosis by inducing intravascular thrombosis leading to infarction (24), and prolonged EGFR activation in cultured cells has been shown to trigger oxidative neuronal injury in adjacent neurons (25). On the other hand, the ability of vessels generated by vascular endothelial growth factor – mediated angiogenesis to perfuse a growing EGFRamplified tumor may reduce the amount of necrosis. T2/T1 and T2 BSC were not predictive of patient survival. Given the recent finding that EGFR genetic changes are not predictive of survival in glioblastoma patients (26, 27), it is not surprising that radiographic variables predictive of EGFR amplification also do not correlate with survival. Whereas a radiographic variable predictive of glioblastoma patient survival would be clinically useful, our objective was to identify variables such as T2/T1 and T2 BSC that were predictive of EGFR amplification. These variables are of value for several reasons. With further analysis, the T2/T1 ratio and T2 BSC may prove to be clinically useful. In a larger data set, a cutpoint T2/T1 or T2 BSC value might be identified through which accurate preoperative predictions of tumor genotype might be possible for some patients. Further studies might also seek radiographic features correlated with expression of the variant III EGFR mutation that is constitutively active with a truncated cytoplasmic domain. EGFR variant III is an independent negative prognostic factor for survival and is overexpressed in 45% of EGFR-amplified and in